AI-Derived Score to Predict Effort Intolerance and Adverse Outcomes Across the Heart Failure Spectrum

To develop and externally validate an artificial intelligence (AI)-driven model to predict effort intolerance (ie, peak oxygen uptake [VO₂] <16 mL/kg/min) in patients at risk for or with established heart failure (HF).

We enrolled a consecutive sample of adults referred for dyspnea or suspected HF. The derivation cohort (Pisa, Italy) included 1333 participants: 351 with reduced (<50%, HFrEF), 371 with preserved (≥50%, HFpEF) left ventricular ejection fraction (LVEF), and 611 with cardiovascular risk factors and/or structural heart disease without overt HF (stages A, B); the external validation cohort (Hasselt, Belgium) included 1101 participants. All participants underwent clinical evaluation, laboratory tests, rest echocardiography, and cardiopulmonary exercise testing.

A neural network including age, sex, body mass index (BMI), hemoglobin, left ventricular systolic mitral annulus tissue velocity (LV S'), systolic pulmonary artery pressure (sPAP), and β-blocker therapy achieved the best discrimination (AUC 0.86 ± 0.01 in derivation; 0.76 ± 0.06 in validation). A simplified 4-variable AI-VO₂ score (BMI, hemoglobin, LV S', sPAP) showed good performance (AUC 0.79 ± 0.05) and independently predicted hospitalization due to HF or all-cause death (adjusted HR 1.06 per point; 95% CI 1.03-1.10) in the derivation cohort. External validation confirmed the predictive and prognostic performance (AUC 0.73 ± 0.02; unadjusted HR 1.18 per point, 95% CI 1.13-1.24). Score-based risk strata (<10, low; 10-13, intermediate; >13, high) showed a significant prognostic gradient (log-rank χ² = 36.8; P < 0.001).

The AI-VO₂ score is a clinically interpretable, externally validated tool for identifying patients with effort intolerance and adverse outcomes across the HF spectrum, supporting personalized risk stratification and management.